Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Kimi K3.1 Pro weights are out: what 1.56 TB buys

Written November 5, 2026

Moonshot published the full weights on November 4 under its own licence. What it takes to run them, and when the API is still the sane route.

What landed

Moonshot launched Kimi K3.1 Pro on October 23, promising the full weights by November 4. The weights are now on Hugging Face as moonshotai/Kimi-K3.1-Pro: 96 safetensors shards totalling 1.56 TB, with no access gate. The model card describes a mixture-of-experts model with 2.8 trillion parameters in total, 104 billion active for each token, 896 experts with 16 selected per token, and a 1,048,576-token context window. Weights are stored at MXFP4 precision with MXFP8 activations, after quantization-aware training. Text and image input are supported. The faster Highspeed variant has no repository of its own; the weights cover K3.1 Pro only.

The model is a post-trained version of Kimi K3, and the card presents it as Moonshot's most capable 3T-class model to date, aimed at agentic coding, browsing, and long-horizon work. The repository was created on November 2 and the weights went public on the promised date.

The licence

The weights ship under the Kimi K3.1 Pro License, Moonshot's own licence rather than a standard open-source one. For most people it works like a permissive licence: use, copy, change, fine-tune, and sell the model, keeping the copyright notice. Two conditions bite at scale. A business that sells access to the model as a service, with group revenue above $20 million over any 12 months, must enter into a separate agreement with Moonshot before commercial use. And a product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must prominently display "Kimi K3.1-Pro" in its interface. Internal use is exempt from both conditions. The licence text matches the Kimi K3 licence apart from the model name, so nothing from the K3 licence carries over by assumption — the K3.1 Pro text is what governs. Read it before building a commercial service on the weights.

What it takes to run

Downloading the model is the easy part. The model card recommends vLLM, SGLang, and TokenSpeed as inference engines, with deployment recipes published for vLLM and SGLang. The simplest vLLM setup is eight data-centre GPUs, such as one NVIDIA B300 node or two GB300 trays. That puts self-hosting out of reach for a single workstation and most small teams: the download alone is 1.56 TB, and the hardware bill dwarfs the API for anything short of heavy, sustained use. No GGUF build was found on Hugging Face on November 4, so the usual single-machine route is not available either. Memory figures circulating for the model are estimates from third parties, not measured runs, so treat any VRAM number as a planning guess until someone publishes a measured setup.

The API is still the sane route for most

On Moonshot's API, K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, with cached input at $0.20, per Moonshot's pricing page. Highspeed doubles the input and output rates to $4 and $20, with cached input staying at $0.20. The model IDs are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. At those rates, renting the model beats self-hosting for intermittent work, prototyping, and any workload that does not run near-constantly on hardware you already own. Self-hosting earns its keep only where the API cannot go: air-gapped use, custom fine-tuning on the weights, or sustained volume large enough that the hardware pays for itself.

The benchmarks, read as vendor claims

Every score below is Moonshot's own, from the model card, and none has been independently rerun on the downloadable files. On Terminal-Bench 4.0 the card reports 65.7%, against 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Artificial Analysis gives the model an intelligence index of 55. The card also reports 92.1 on BrowseComp and 70.0 on HLE-Full with tools. Those numbers make K3.1 Pro competitive with the frontier on the tests Moonshot chose; they do not show how it behaves on a specific codebase, where the harness, the effort setting, and the surrounding agent matter as much as the model. The card's own footnotes say its scores use maximum reasoning effort, so a cheaper effort setting will score lower.

What to do

  1. Decide rent or run first. If the work is intermittent, use the API. Download the weights only with eight data-centre GPUs available and a workload that justifies them.
  2. Read the licence before commercial use. The $20 million revenue condition and the display requirement apply to services and large products, not to internal use — but check which side of the line your use falls on.
  3. Pick the engine from the card. vLLM, SGLang, and TokenSpeed are the recommended engines, with published recipes for the first two. Do not assume a GGUF build exists.
  4. Test on your own work before switching. The vendor benchmarks are a starting point. Run the model on your own tasks, at the effort setting you will actually pay for, and compare cost per finished task rather than cost per token.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Run LongCat-2.5 in Claude Code without the 200K cap · What an agent costs, in plain figures · Mistral Large 4 Le Chonk: a trillion parameters, open